fix(openai_compat): fix tool call serialization for strict OpenAI-compatible providers

[securityguy]

Summary

• Fix empty content with tool_calls: When the assistant responds with only tool calls (no text), the serialized message previously included "content": "". The OpenAI spec allows content to be absent when tool_calls is present, and some strict providers (e.g. certain backends routed via OpenRouter) reject the empty string. The field is now omitted in that case.

• Add strict_compat option: Some providers reject non-standard fields that picoclaw preserves for extended model features (reasoning_content, extra_content, thought_signature). A new per-model strict_compat: true config option strips these fields before serialization. Defaults to false so existing behaviour is unchanged.

  Also refactors the HTTPProvider constructors from a growing list of named variants into a single NewHTTPProviderWithOptions(apiKey, apiBase, proxy string, opts ...openai_compat.Option), making future options trivial to add.

Usage

{
  "model_name": "openrouter-strict",
  "model": "openrouter/anthropic/claude-sonnet-4-5",
  "api_key": "sk-or-v1-...",
  "strict_compat": true
}

Test plan

• Existing openai_compat tests pass
• TestSerializeMessages_OmitsContentWhenEmptyAndToolCallsPresent — content absent when tool_calls present and content is empty
• TestSerializeMessages_IncludesContentWhenNonEmptyWithToolCalls — content preserved when non-empty alongside tool_calls
• TestSerializeMessages_StrictCompat_StripsReasoningContent — reasoning_content stripped when strict_compat=true
• TestSerializeMessages_StrictCompat_StripsExtraContent — extra_content stripped when strict_compat=true
• TestSerializeMessages_StrictCompat_StripsThoughtSignature — thought_signature stripped when strict_compat=true
• TestSerializeMessages_NoStrictCompat_PreservesFields — all fields preserved when strict_compat=false

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Copilot]

Pull request overview

This PR improves compatibility with strict OpenAI-compatible chat completion endpoints and expands provider/channel infrastructure to better support multi-provider fallback, CLI-based providers, multi-bot Telegram setups, and more robust cron/gateway behaviors.

Changes:

• Fix OpenAI-compatible message serialization (omit empty content when tool_calls exist; add strict_compat to strip non-standard fields).
• Add per-candidate provider dispatch (dispatcher + cached per-model providers) and extend CLI providers (Gemini CLI, request timeouts, improved tool-call extraction).
• Expand operational features: multi-bot Telegram support, cron store reload semantics + outbound publishing, and launcher gateway health probing.

Reviewed changes

Copilot reviewed 36 out of 36 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
web/backend/api/model_status.go	Treats gemini-cli as a CLI protocol needing runtime probe/availability handling.
web/backend/api/gateway.go	Avoids auto-starting gateway when an externally managed gateway is already healthy; improves status reporting.
pkg/tools/toolloop.go	Adds optional dispatcher/candidate/fallback execution path for tool loops.
pkg/tools/subagent_tool_test.go	Updates tests to new SubagentManagerConfig constructor API.
pkg/tools/subagent.go	Refactors subagent manager construction/config; adds per-agent dispatch and response attribution.
pkg/tools/spawn_test.go	Updates tests to new SubagentManagerConfig constructor API.
pkg/tools/spawn.go	Applies allowlist checks to self-spawn cases (using caller agent ID).
pkg/tools/cron.go	Publishes cron job responses directly to outbound bus (deliver, command output, and agent responses).
pkg/providers/tool_call_extract_test.go	Adds comprehensive tests for extracting/stripping tool_calls JSON (pretty JSON, code fences, mixed text).
pkg/providers/tool_call_extract.go	Reworks CLI tool-call extraction/stripping (code-fence tolerant, args as string/object).
pkg/providers/openai_compat/provider_test.go	Adds tests for empty-content-with-toolcalls and strict-compat field stripping.
pkg/providers/openai_compat/provider.go	Implements strictCompat option and new serialization that omits empty content when tool_calls exist.
pkg/providers/http_provider.go	Simplifies HTTPProvider constructors into NewHTTPProviderWithOptions.
pkg/providers/gemini_cli_provider_test.go	Adds tests for Gemini CLI provider prompt building, response parsing, and factory integration.
pkg/providers/gemini_cli_provider.go	Introduces Gemini CLI provider implementation (subprocess, JSON output parsing, usage aggregation).
pkg/providers/factory_provider.go	Wires strict_compat + request_timeout options into HTTP providers; adds gemini-cli; adds CLI timeouts for claude/codex.
pkg/providers/error_classifier.go	Uses errors.Is for deadline exceeded classification (better wrapped-error handling).
pkg/providers/dispatch_test.go	Adds dispatcher cache/flush/thread-safety tests.
pkg/providers/dispatch.go	Adds ProviderDispatcher to create/cache per protocol/model providers.
pkg/providers/codex_cli_provider.go	Adds request timeout support and clearer timeout errors for codex CLI.
pkg/providers/claude_cli_provider_test.go	Updates tests to ensure system prompt is passed via stdin, not argv; adds stdin prompt tests; removes brace-matcher tests.
pkg/providers/claude_cli_provider.go	Moves system prompt into stdin payload; adds request timeout; improves error reporting (stdout/stderr surfaced).
pkg/cron/service.go	Reloads cron store when file changes externally; avoids saving when no due jobs; tracks store file mtime.
pkg/config/config.go	Adds strict_compat model option; introduces telegram_bots config + normalization from legacy channels.telegram.
pkg/channels/telegram/telegram.go	Supports constructing Telegram channels from explicit TelegramConfig + channel name (for named bots).
pkg/channels/telegram/init.go	Registers Telegram bot factory for named telegram bots initialization.
pkg/channels/registry.go	Adds global registration/accessors for a TelegramBotFactory (avoids import cycles).
pkg/channels/manager.go	Initializes multiple named Telegram bots; factors out dependency injection; inherits rate limits for channel variants.
pkg/agent/loop_test.go	Updates AgentLoop construction to accept dispatcher parameter.
pkg/agent/loop.go	Adds dispatcher usage for fallback candidates; flushes dispatcher on reload; tweaks direct-processing peer metadata; publishes assistant text when tool_calls present.
config/config.example.json	Documents bindings + telegram_bots usage in example config.
cmd/picoclaw/internal/gateway/helpers.go	Instantiates and passes ProviderDispatcher into AgentLoop.
cmd/picoclaw/internal/cron/helpers.go	Improves cron list output formatting and truncates message display.
cmd/picoclaw/internal/agent/helpers.go	Instantiates and passes ProviderDispatcher into AgentLoop.
README.md	Replaces README with fork status + configuration guide for added features.
.github/dependabot.yml	Removes dependabot configuration.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Same issue for the non-fallback dispatch path: when Dispatcher.Get(first.Provider, first.Model) fails, falling back to config.Provider.Chat(first.Model) can send the request to an unintended provider. Better to return the dispatcher error (or explicitly document/guard this behavior) so a missing dispatcher entry doesn't silently change routing.

[Copilot]

ProcessDirectWithChannel unconditionally sets msg.Peer.Kind to "channel" for any non-"direct" chatID. That can break peer-based bindings that expect "direct" or "group" (e.g., bindings matching Telegram DMs/groups), and it may misroute cron/CLI-originated messages compared to real inbound messages. Consider either (a) not setting Peer unless the caller can supply the correct kind, or (b) extending the API / cron payload to carry the peer kind captured from the original inbound message.

Suggested change

	// Set peer so channel-based bindings (e.g. a specific Slack channel mapped
	// to a named agent) are matched by the route resolver, exactly as they are
	// for live inbound messages.
	if chatID != "" && chatID != "direct" {
	msg.Peer = bus.Peer{Kind: "channel", ID: chatID}
	}

[Copilot]

Typo: "This repos" should be "This repo".

Suggested change

	CAUTION: This repos is not maintained and should not be used. It has been left here in case the branches are useful to the picoclaw maintainers.
	CAUTION: This repo is not maintained and should not be used. It has been left here in case the branches are useful to the picoclaw maintainers.

[Copilot]

The note says legacy channels.telegram is normalized to a telegram-default channel, but the code path (TelegramBotConfig.ChannelName) maps id "default" to the channel name "telegram". Update the README to match the actual channel name used for backward compatibility.

Suggested change

	**Note:** The legacy single-bot `channels.telegram` config is still supported and is automatically normalized to a `telegram-default` channel for backward compatibility.
	**Note:** The legacy single-bot `channels.telegram` config is still supported and is automatically normalized to a `telegram` channel for backward compatibility.

[Copilot]

PR title/description focuses on openai_compat serialization/strict_compat, but this PR also includes broad changes (providers dispatching, telegram multi-bot support, cron behavior, gateway probing, README rewrite, etc.). Either split into focused PRs or expand the PR description to cover the additional behavioral changes and their test plan, to make review and risk assessment tractable.

[Copilot]

stripToolCallsFromText validates tool_calls JSON using a fence-stripped copy, but then removes only the {...} range from the original text. When the tool_calls blob is inside markdown code fences, this leaves dangling /json fences in the returned content (so the caller may treat it as non-empty user-visible text). Consider detecting fenced blocks and removing the entire fenced region, or compute removal indexes against the fence-stripped string and map them back correctly.

Suggested change

	// Use positions in the original (pre-fence-stripped) text to preserve
	// any content outside code fences. If the text was fence-wrapped we strip
	// the whole thing; otherwise cut out just the JSON range.
	// If the tool_calls JSON is inside a markdown code fence, remove the
	// entire fenced block so we don't leave dangling ```/```json markers.
	if fenceStart := strings.Index(text, "```"); fenceStart != -1 {
	if fenceEndRel := strings.Index(text[fenceStart+3:], "```"); fenceEndRel != -1 {
	fenceEnd := fenceStart + 3 + fenceEndRel
	// Remove the whole fenced region, preserving any surrounding text.
	return strings.TrimSpace(text[:fenceStart] + text[fenceEnd+3:])
	}
	}
	// No fences detected; fall back to removing just the JSON range in the
	// original text, preserving any surrounding content.

[Copilot]

In the per-candidate fallback path, if Dispatcher.Get(providerName, model) fails, this code silently falls back to config.Provider.Chat using the candidate's model. That can route the request to the wrong provider/base URL and can also mask a misconfigured candidate (it should be treated as a failed attempt so the fallback chain can proceed to the next candidate). Consider returning the dispatcher error instead (or wrapping it) so this attempt fails cleanly.

[sipeed-bot]

@securityguy Hi! This PR has been inactive for over a week. If there's no update in the next 7 days, it will be closed automatically. If you're still working on it, just leave a comment to keep it open!